Floristic Evidence for Alternative Biome States in Tropical Africa

Floristic evidence for alternative biome states in tropical Africa

J. C. Alemana,b, A. Fayollea,1, C. Favierc, A. C. Staverd, K. G. Dextere,f, C. M. Ryane, A. F. Azihoug, D. Baumanh,i, M. te Beestj,k,l, E. N. Chidumayom, J. A. Comiskeyn,o, J. P. G. M. Cromsigtj,k,p, H. Dessardq,r, J.-L. Douceta, M. Finckhs, J.-F. Gilleta, S. Gourlet-Fleuryq,r, G. P. Hempsont, R. M. Holdou, B. Kirundav, F. N. Kouamew, G. Mahya, F. Maiato P. Gonçalvesx, I. McNicole, P. Nieto Quintanoe, A. J. Plumptrev,y,z, R. C. Pritcharde,aa, R. Revermanns,bb, C. B. Schmittcc,dd, A. M. Swemmeree, H. Talilaff, E. Woollene, and M. D. Swainegg

aGembloux Agro-Bio Tech, Université de Liège, 5030 Gembloux, Belgium; bDépartement de Géographie, Université de Montréal, Montréal, QC H2V 0B3, Canada; cInstitut des Sciences de l’Evolution–Montpellier, CNRS, Ecole Pratique des Hautes Etudes, Institut de Recherche pour le Développement, Université de Montpellier, 34000 Montpellier, France; dDepartment of Ecology and Evolutionary Biology, Yale University, New Haven, CT 06520; eSchool of GeoSciences, University of Edinburgh, EH8 9YL Edinburgh, United Kingdom; fTropical Diversity Section, Royal Botanic Garden Edinburgh, EH3 5LR Edinburgh, United Kingdom; gLaboratory of Applied Ecology, Faculty of Agronomic Sciences, University of Abomey-Calavi, 01 BP 526 Cotonou, Benin; hEnvironmental Change Institute, School of Geography and the Environment, University of Oxford, OX1 3QY Oxford, United Kingdom; iPlant Ecology and Biogeochemistry, Université Libre de Bruxelles, 1050 Bruxelles, Belgium; jCopernicus Institute of Sustainable Development, Utrecht University, 3584 CS Utrecht, The Netherlands; kCentre for African Conservation Ecology, Nelson Mandela University, 6031 Port Elizabeth, South Africa; lGrasslands-Forests-Wetlands Node, South African Environmental Observation Network, 3201 Pietermaritzburg, South Africa; mMakeni Savanna Research Project, Ridgeway, 1001 Lusaka, Zambia; nInventory and Monitoring Program, National Park Service, Fredericksburg, VA 22405; oSmithsonian Institution, Washington, DC 20002; pDepartment of Wildlife, Fish, and Environmental Studies, Swedish University of Agricultural Sciences, SE-901 83 Umeå, Sweden; qForêts et Sociétés, Centre de Coopération Internationale en Recherche Agronomique Pour le Développement, Université de Montpellier, 34000 Montpellier, France; rCentre de Coopération Internationale en Recherche Agronomique Pour le Développement, Forêts et Sociétés, 34398 Montpellier, France; sBiodiversity, Evolution and Ecology of Plants, Institute of Plant Science and Microbiology, University of Hamburg, 22609 Hamburg, Germany; tCentre for African Ecology, School of Animal, Plant and Environmental Sciences, University of the Witwatersrand, 2000 Johannesburg, South Africa; uOdum School of Ecology, University of Georgia, Athens, GA 30602; vWildlife Conservation Society, Kampala, Uganda; wNature Sciences Unit, University Nangui Abrogoua, 31 BP 165 Abidjan, Côte d’Ivoire; xHerbário do Lubango, Instituto Superior de Ciências da Educação da Huíla, C.P. 230 Lubango, Angola; yHead Key Biodiversity Area Secretariat, c/c BirdLife International, CB2 3QZ Cambridge, United Kingdom; zConservation Science Group, Zoology Department, Cambridge University, CB2 3EJ Cambridge, United Kingdom; aaGlobal Development Institute, University of Manchester, M13 9PL Manchester, United bb cc Kingdom; Faculty of Natural Resources and Spatial Sciences, Namibia University of Science and Technology, Windhoek, Namibia; Center for ENVIRONMENTAL SCIENCES Development Research, University of Bonn, 53113 Bonn, Germany; ddNature Conservation and Landscape Ecology, University of Freiburg, 79106 Freiburg, Germany; eeSouth African Environmental Observation Network, SAEON Ndlovu Node, Phalaborwa, 1390, South Africa; ffDepartment of Ecotourism and Biodiversity Conservation, College of Agriculture and Natural Resource, Madda Walabu University, Bale Robe, Ethiopia; and ggInstitute of Biological and Environmental Sciences, University of Aberdeen, AB24 3FX Aberdeen, United Kingdom

Edited by Robert John Scholes, University of the Witwatersrand, Wits, South Africa, and approved September 7, 2020 (received for review June 8, 2020)

The idea that tropical forest and savanna are alternative states is crucial to how we manage these biomes and predict their future Significance under global change. Large-scale empirical evidence for alternative stable states is limited, however, and comes mostly from We develop a biogeographic approach to analyzing the pres- the multimodal distribution of structural aspects of vegetation. ence of alternative stable states in tropical biomes. Whilst These approaches have been criticized, as structure alone cannot forest–savanna bistability has been widely hypothesized and separate out wetter savannas from drier forests for example, and modeled, empirical evidence has remained scarce and contro- there are also technical challenges to mapping vegetation structure versial, and here, applying our method to Africa, we provide in unbiased ways. Here, we develop an alternative approach to de- large-scale evidence that there are alternative states in tree limit the climatic envelope of the two biomes in Africa using tree species composition of tropical vegetation. Furthermore, our species lists gathered for a large number of forest and savanna sites results have produced more accurate maps of the forest and distributed across the continent. Our analyses confirm extensive cli- savanna distributions in Africa, which take into account dif- matic overlap of forest and savanna, supporting the alternative sta- ferences in tree species composition, and a complex suite of ble states hypothesis for Africa, and this result is corroborated by determinants. This result is not only important for under- paleoecological evidence. Further, we find the two biomes to have standing the biogeography of the continent but also, to guide highly divergent tree species compositions and to represent alterna- large-scaled tree planting and restoration efforts planned for tive compositional states. This allowed us to classify tree species as the region. forest vs. savanna specialists, with some generalist species that span both biomes. In conjunction with georeferenced herbarium records, Author contributions: J.C.A., A.F., and M.D.S. designed research; J.C.A., A.F., A.F.A., we mapped the forest and savanna distributions across Africa and D.B., M.t.B., E.N.C., J.A.C., J.P.G.M.C., H.D., J.-L.D., M.F., J.-F.G., S.G.-F., G.P.H., R.M.H., quantified their environmental limits, which are primarily related to B.K., F.N.K., G.M., F.M.P.G., I.M., P.N.Q., A.J.P., R.C.P., R.R., C.B.S., A.M.S., H.T., E.W., and M.D.S. performed research; J.C.A., A.F., and C.F. contributed new reagents/ precipitation and seasonality, with a secondary contribution of fire. analytic tools; J.C.A. and A.F. analyzed data; A.C.S., K.G.D., C.M.R., and D.B. com- These results are important for the ongoing efforts to restore African mented on the paper; A.F.A., M.t.B., E.N.C., J.A.C., H.D., J.-L.D., J.-F.G., B.K., F.N.K., ecosystems, which depend on accurate biome maps to set appropri- G.M., F.M.P.G., I.M., P.N.Q., R.C.P., R.R., H.T., E.W., and M.D.S. contributed to data; ate targets for the restored states but also provide empirical evidence J.P.G.M.C., M.F., S.G.-F., G.P.H., R.M.H., A.J.P., C.B.S., and A.M.S. contributed to data and commented on the paper; and J.C.A., A.F., A.C.S., K.G.D., C.M.R., and M.D.S. wrote for broad-scale bistability. the paper. The authors declare no competing interest. alternative stable states | tropical biomes | tree species composition | This article is a PNAS Direct Submission. precipitation and seasonality | fire Published under the PNAS license. 1To whom correspondence may be addressed. Email: [email protected]. ree cover and canopy openness are commonly used to dif- This article contains supporting information online at https://www.pnas.org/lookup/suppl/ Tferentiate tropical forests and savannas, but the difference doi:10.1073/pnas.2011515117/-/DCSupplemental. between the two biomes is not just a matter of structure (1).

www.pnas.org/cgi/doi/10.1073/pnas.2011515117 PNAS Latest Articles | 1of8 Downloaded by guest on September 27, 2021 Whereas tropical moist forests form a closed canopy with a used georeferenced herbarium collections (24, 25) to expand the complex vertical structure, savannas are more open, allowing floristic information spatially, to describe the complete distri- fire- and drought-adapted grasses to grow in the understory. bution of forest and savanna across Africa, and to identify the Typically, forest dominates in wetter areas, while savannas occur determinants of their present day distribution. in drier, seasonal areas (1, 2), although transitions between forest and savanna are not rigidly determined by climate (2–4). Soils Extensive Climatic Overlap of Forest and Savanna in Africa and topography can be locally and regionally important, but at Alongside the expected patterns of forest in wetter areas and intermediate rainfall (between 1,000 and 2,500 mm globally), savanna in drier and more seasonal ones (1, 2), also retrieved forest and savanna, both widespread, potentially represent al- here (Fig. 1A and SI Appendix, Fig. S1), we found an extensive ternative stable states maintained by feedbacks between tree climatic area (Fig. 1 B and C and SI Appendix, Fig. S2) within cover and disturbances—specifically fire (3, 5–7) and chronic which both forest and savanna sites are widespread, which we herbivory (8). As a result, forest and savanna tree species show term the “bistable” region. Mean annual precipitation (MAP) contrasting adaptations (9, 10), and transitions across the partly differentiates the forest, the bistable region, and the sa- forest–savanna boundary are characterized by high species vanna, although the climatic gradient used to determine and map turnover (10, 11). them is more complex and integrates precipitation and season- Despite these functional and floristic differences, most recent ality (Materials and Methods has the details of the principal efforts devoted to understanding forest–savanna transitions have component analysis [PCA] on gridded climate data). We find a relied solely on information about the canopy structure, whether large extent of forest, 1.8 million km2, covering almost the whole using satellite-based maps of percent tree cover (3, 5, 6, 12), Guineo-Congolian region (Fig. 1D), in line with the stability of canopy height (13), or using field data on tree basal area (14). forest in Lower Guinea suggested over the last two millennia by However, these studies have two types of shortcomings. The first pollen records (29) but challenging previous findings of bist- type is linked to the use of remote sensing products, which are ill ability across the entire Congo basin (3). Savanna was found to suited to capture the structural difference between savanna and occupy the majority of the areas north, east, and south of the forest. Most tree cover products are parameterized to identify Congo basin, under drier and more seasonal conditions and to- trees greater than 5 m in height (15, 16), and their spatial res- taling almost 8.4 million km2. The savanna notably includes the olution is not sufficient to correctly characterize vegetation west coast of southern Africa (Fig. 1D), which contrasts with structure. Even though Landsat-based tree cover is available at description of this area as thicket, scrub, or bush land (28). This 30-m resolution and Moderate Resolution Imaging Spectror- is due to the presence of northern (Sahel) and eastern (Horn) adiometer Active Fire (MODIS) tree cover at 250 m, the accu- savanna sites with extremely arid climate in the floristic surveys, racy of these products is low in open systems (16), and although which extended the climatic envelope of savanna this far, and it increases with spatial aggregation, it then fails to represent the also because northern and western savannas are not completely fine-scale heterogeneity characteristic of savanna. Moreover, the analogous climatically to southern and eastern ones (27). tree cover threshold used to differentiate forest from savanna Meanwhile, the bistable region covers a vast area of more than seems to depend strongly on calibration, with higher thresholds 7.5 million km2, often described as woodland (28), and spans a (75%) for Landsat tree cover (12) and lower (55 to 65%) for broad MAP range (700 to 1,900 mm) (SI Appendix, Fig. S2). The MODIS tree cover (3, 17). The second type of shortcomings is bistable region was previously proposed to span a 1,000- to even more important since defining forest and savanna using 2,000-mm MAP range in Africa (6) or alternately, to extend to only canopy structure misses key aspects of forest vs. savanna drier sites (>650 mm) in a study restricted to savannas (30) and function tied to differences in the structure of the grass layer and to wetter sites (1,000 to 2,500 mm) in the global tropics (3, 12). in the tree species composition (1). The relatively open canopies Our bistable region can be separated into two subregions: a of drier forests can be structurally similar to those of wetter sa- wetter region more likely to be forest (hereafter, bistable forest) vannas (18) but are not separable with a structural approach, and a drier and more seasonal one more likely to be savanna even though the two differ dramatically in terms of species (hereafter, bistable savanna). Bistable forest is located in the composition and ecosystem function (19). Also, systems with periphery of the Guineo-Congolian region but also includes the similar physiognomies can have different histories, reflecting Congolese Batéké Plateau and the Dahomey gap, which are human land use practices and recent disturbances (12). For in- currently dominated by savanna. This stresses the importance of stance, forests subjected to human disturbances (such as clear- factors other than climate in the outcome of forest–savanna ing) can appear more like savannas, while some practices bistability. The Batéké Plateau corresponds to the northern limit (particularly fire suppression) can make savannas appear more of the hyper–well-drained relict dunes of the Kalahari sands, like forests (19, 20). while the Dahomey gap is well known to have shifted to savanna Biodiversity data available for sub-Saharan Africa (21–23) and between 4,500 and 3,400 y ago following an abrupt climatic for tropical Africa (24–27) offer new opportunities for differ- change (31). Thus, the predominance of savanna in areas more entiating and mapping the tropical forest and savanna biomes at likely to be forest has arisen because of either soil characteristics continental scale. Here, we delineate the distribution of forest (i.e., sandy soils) (30) or the legacy of past climatic changes (29) and savanna and test for forest–savanna bistability using a bio- and/or past disturbances (31, 32). geographic approach, which reflects the evolutionary history and Long-term paleoecological data (from lacustrine and soil ecology of each biome and does not rely on any structural met- cores) around the Gulf of Guinea largely corroborate these areas rics. We combined data on native tree species for sites identified of biome transition and stability (Fig. 1 D, Inset). For the bistable earlier as forest (26) (n = 455) and savanna (27) (n = 298) and region, forest-to-savanna transitions were predominant with 16 covering the full extent of these biomes in Africa (28). Each site transitional vs. 9 stable sites, while for the forest, stable biome consists of a species list assigned to either forest or savanna by identity was retrieved for 10 of 13 paleosites. For the savanna, the original authors or by the data contributors who conducted the stability of savanna vegetation has been demonstrated the floristic surveys according to vegetation structure, species around Lake Tilla (#21 on the map) in Nigeria (33). composition, and/or ecosystem functioning (Materials and Current fire regime is also an important determinant of sa- Methods). We used tree species composition from these 753 vanna distribution worldwide (2, 3, 6), specifically in relatively floristic surveys to determine the climatic envelopes of the forest wetter areas (20, 30), and here, we found that bistable savanna is and savanna biomes (and their overlap) and to analyze the dif- more likely to burn than either forest or bistable forest and also, ferences in tree species composition between them. We then than savanna (SI Appendix, Fig. S3). The latter is explained by

2of8 | www.pnas.org/cgi/doi/10.1073/pnas.2011515117 Aleman et al. Downloaded by guest on September 27, 2021 ENVIRONMENTAL SCIENCES

Fig. 1. Climatic envelope of the forest and savanna biomes in Africa. To determine the climatic envelopes of the two tropical biomes, the dominant climatic gradients were first identified with a PCA of gridded climatic variables. (A) Each point corresponds to the center of a 0.5° pixel, with pixels containing floristic surveys in forest (green) and savanna (orange) sites indicated. Red and blue arrows indicate the influence of temperature and precipitation variables, re- spectively. (B) Distribution of forest (green) and savanna (orange) sites along an axis of MAP (in millimeters). (C) Frequency distribution of forest (green) and savanna (orange) sites along a complex precipitation and seasonality gradient (PC1; A), with the climatic area where savanna and forest both occur plotted in light orange (where savanna is more common) and light green (where forest is more common). The dashed line corresponds to an equal probability of savanna and forest. (D) Map of forest and bistable forest along with bistable savanna and savanna, with the locations of floristic surveys in forest (green) and savanna (orange) sites. White pixels are outside the geographic extent and/or the environmental range covered by the floristic surveys. The latter was defined by a convex hull on the site scores on PC1 and PC2 (dashed line in A). Major water bodies and rivers are shown in blue. The paleoecological data available around the Gulf of Guinea retrieved from published records of lacustrine fossil pollen (n = 24), lacustrine phytoliths (n = 1), phytoliths (n = 4), and δ13Cofsoil organic matter (n = 14) from soil profiles are shown in Inset. The paleosites provided information on stable (black) and transitional (gray) sites during the Holocene (Dataset S2 has the code correspondence of paleosites).

Aleman et al. PNAS Latest Articles | 3of8 Downloaded by guest on September 27, 2021 the lower productivity of drier savannas (30) but could also be mainly of savanna specialists. Indeed, forest specialists were linked to the presence of semidesert grasslands (28) in our study found to dominate the Guineo-Congolian region and its pe- area (Sahel, Namib, Kalahari), where the discontinuous grass riphery and to be abundant in East African Mountains (Fig. 2A). layer prevent the spread of fires. Within the Guineo-Congolian region, Upper and Lower Guinea were better sampled than Congolia, for which herbarium records African Forests and Savannas Have Distinct Floristic are sparse, and Upper Guinea includes the warm and wet forest Composition sheltered by the Guinean Highlands in Sierra Leone and Guinea We found a marked dissimilarity in tree species composition (Conakry). The latter were not included in our mapped extent of between forest and savanna sites, based on floristic surveys. forest (Fig. 1D) likely because the floristic surveys we used do Pairwise comparisons showed a mean species turnover of 98% not cover this climate zone (positive scores on the first two between forest and savanna sites, much higher than the turnover principal components of climate, PC1 and PC2) (Fig. 1A). Sa- within the same biome (SI Appendix, Fig. S4). This result con- vanna specialists dominate in the north and west of Africa, in- firms the divergence of the forest and savanna floras (22) and is cluding the Dahomey gap, and in the east and south (Fig. 2C). consistent with the results of plot-based studies of forest–savanna Interestingly, generalist species were found to be widespread and transition in South Africa (10), Ghana (11), and across the tro- present within each region, and while they are more frequent in pics (18). Because forest and savanna represent alternative the bistable region, the difference is small (SI Appendix, Fig. compositional states, we were confident in categorizing tree S5B), and generalists almost never dominate (Fig. 2B and SI species into significant indicators of forest or forest specialists Appendix, Fig. S6). (n = 825 species, 48%) and significant indicators of savanna or savanna specialists (n = 523, 31%). Species that were not a sig- Savanna–Forest Coexistence Is Possible but Restricted nificant indicator of either forest or savanna were interpreted as Spatially generalists (n = 359, 21%). Our group of generalists included Based on a biome index, indicating the relative dominance of true generalists but also, rare species or those otherwise rarely savanna (−1) and forest (1) specialists (Materials and Methods), represented in the floristic surveys (Fig. 2). Evidence for biome we find that forest–savanna coexistence is limited (Fig. 3), even specialization was even stronger when only common species were though the climate envelope where it is possible is large (42.4% analyzed (i.e., those present at ≥10 sites), with only 8% of of the study area and 25% of the African continent) (Fig. 1D). common species classified as generalists. It is worth noting that Vast and continuous areas are dominated by the extremes of the our results for 1,707 tree species represent only a fraction of the biome index, representing either strict savanna or strict forest (SI circa 45,000 flowering plant species reported for sub-Saharan Appendix, Fig. S7), and these extremes of the biome index cor- Africa (34). The specificity of the forest and the savanna flora respond to recognized centers of endemism (28) (i.e., the is therefore likely underestimated since grasses and forbs were Guineo-Congolian region where forests are predominant and the not included, despite being a highly distinctive component of the Sudanian and Zambezian regions where savannas and wood- savanna flora (22). Functionally, however, these results are de- lands are predominant, and both have specific floras and faunas) monstrative since the tree flora already captures some functional (21–24). In contrast, intermediate values of the biome index, differences between the forest and savanna biomes (10), but it which correspond to the third mode in the frequency distribution would be of great interest to extend the analyses to other growth (Fig. 3, Inset), appear spatially restricted (SI Appendix, Fig. S7). forms to test the patterns observed for trees. The zones of intermediate biome index, which arise either from The regions described above based on climate (Fig. 1) have frequent generalists (as in East Africa) or from a mixture of distinct compositions (SI Appendix, Fig. S5), assessed using an forest and savanna specialists elsewhere (Fig. 2B and SI Ap- independent occurrence dataset (Fig. 2 and Materials and pendix, Fig. S6), correspond to long recognized transition zones, Methods) derived from georeferenced herbarium records (24, forest– savanna mosaics (28), and to recently deforested areas 25). Forest and bistable forest are composed mainly of forest (35). A deeper investigation shows, however, that the possible specialists, and savanna and bistable savanna are composed intermediate state mostly corresponds to data-deficient areas

Fig. 2. Distribution of forest specialist, generalist, and savanna specialist tree species. To test for specialization toward the forest and savanna biomesby individual tree species, we applied the IndVal procedure (50) to the presence matrix of the 1,707 species in the 753 sites (455 forest and 298 savanna sites) and obtained a classification of species into forest specialists (n = 825 species) and savanna specialists (n = 523), with nonsignificant indicators resulting in species being interpreted as generalists (n = 359). For each 0.5° pixel containing herbarium records for at least 5 of our 1,707 species, we computed and mapped the percentage of (A) forest specialists, (B) generalists, and (C) savanna specialists. White pixels thus correspond to a paucity of georeferenced herbarium records for our classified tree species. Major water bodies are shown in blue. The frequency distribution of the number of sites in which (A) forest specialists, (B) generalists, and (C) savanna specialists occurred in the original floristic surveys is given in Inset of each panel, illustrating the higher frequency of extremely infrequent species in the group of generalist species, in comparison with the specialists.

4of8 | www.pnas.org/cgi/doi/10.1073/pnas.2011515117 Aleman et al. Downloaded by guest on September 27, 2021 data (Fig. 4A) and than the raw data (Figs. 2 and 3). The latter possibly results from preferential sampling of forest trees in herbarium collections. More importantly, our results support the primary role of precipitation (1, 2), precipitation seasonality (36), and fire (2, 3, 6, 30) in the distribution of both forest and savanna in Africa, while most analyses of the determinants of savanna distribution have so far ignored forest (2, 30). There- fore, the predictions of the random forest (Fig. 4B) should be considered as our best estimate of the current distribution of the forest and savanna biomes across the African continent. Conclusions and Practical Implications In this study, we provided evidence for two compositional states across the African continent, alternatively dominated by forest and savanna species. These two states can coexist in a wide range of climates, but our biome index showed that current day spatial coexistence is spatially restricted, with one state or the other dominating over vast areas. This supports the notion that the two biomes are stable alternatives and consistent with the predictions of alternative stable state theory, suggests the potential for Fig. 3. Spatial distribution of the biome index across Africa. The distributions of forest specialists, generalists, and savanna specialists derived from abrupt shifts in composition under external perturbations such as georeferenced herbarium records were used to devise a biome index based climate change, or altered fire regimes, as demonstrated by pa- on tree species composition and computed at 0.5° resolution. The biome leoecological evidence. With our correlative approach, we were index tracks the biome specialization of each pixel, with values toward −1 not able to determine the ecological mechanisms allowing sa- representing the dominance of savanna specialists and values toward +1 vanna persistence, but the facts that fire is more frequent in the representing the dominance of forest specialists. White pixels correspond to bistable savanna area and that fire is the second determinant of a lack of georeferenced herbarium records for our species. Major water the biome index after the precipitation and seasonality gradient bodies are shown in blue. The frequency distribution of the biome index is support the hypothesis of fire acting as a major feedback given in Inset. mechanism, allowing savanna under climates favorable to forest. ENVIRONMENTAL SCIENCES These results are important for the ongoing efforts to restore African ecosystems, which depend on accurate biome maps to since the trimodal signal shifted toward a bimodal signal when set appropriate targets for the restored states. This will help including only pixels with at least 5 or 10 of our classified tree avoid, for example, inappropriate planting of forest trees or at- species (SI Appendix, Fig. S8). Thus, the floristic surveys (26, 27) tempts to “restore” forest in savanna areas. Up to now, because combined with georeferenced herbarium records (24, 25) suggest of the wide climatic zone where forest and savanna are both that there are two floristic states in the intact tropical African widespread, savanna has often been mischaracterized as de- vegetation. graded forest (28, 37) and has been seen as a target for higher Finally, we provide estimates of the current distribution of the carbon storage via afforestation (37, 38). This viewpoint neglects forest and savanna biomes across the African continent com- several features of savannas, including their substantial below- bining both floristic and environmental information (Fig. 4). ground carbon storage (39), their high biodiversity and ende- First, we recomputed the biome index on interpolated species mism (40), and their socioeconomic value (41). Our biome index distributions derived from ordinary kriging (SI Appendix, Fig. S9) map therefore is a useful tool for restoration, as it could help set while accounting for spatial autocorrelation (Fig. 4 A, Inset). This appropriate species targets, and to identify degraded forest and shows that there is a strong spatial structure for specialist encroached savanna areas, which should be restored using forest species—whether forest or savanna—while generalist species do and savanna specialists, if needed. Indeed, the current trend not show any spatial structure, suggesting that they do not cor- within bistable savannas is toward woody encroachment (42), respond to an ecologically meaningful third group. We addi- and in this case, restoration strategies mainly involve tree cutting tionally include principle components of climate (PC1 and PC2 and prescribed fire to restore biodiversity and ecosystem services from Fig. 1) and soil information (sand percentage and cation (43). Conversely, restoring arid savannas to prevent desertifica- exchange capacity), as well as data on fire and herbivory (Ma- tion might involve tree and shrub planting (44), and in this case, terials and Methods has the source of the datasets) as potential using savanna specialists is and has been warranted. In the spa- determinants of the biome index in a random forest regression tial extent where both biomes currently coexist, potentially both (Fig. 4B). Topographic and hydrologic factors (e.g., seasonal may be restored, but a deeper investigation is needed (for ex- flooding, topographically mediated frost occurrence, or toxic ample, using long-term data) to determine the baseline biome heavy metal concentrations in soils) that are important deter- (45), and human preference should also be considered. minants of mosaic landscapes locally (28) were not included at this continental scale. The predictions of our random forest Materials and Methods model on a validation subset of 10% of data pixels were accurate To delineate the distribution of forest and savanna across the African con- 2 (R = 0.81 with a nonsignificant intercept and a predicted vs. tinent and to identify their underlying determinants, we developed a bio- observed slope of 0.82 ± 0.03). This analysis of the joint deter- geographic approach consisting of four steps, each testing a specific minants of both forest and savanna showed that precipitation hypothesis. First, we used the location of specific sites originally classified as and seasonality (PC1) are the main determinants of the biome forest (26) or savanna (27) to test the forest–savanna bistability [i.e., whether index. Fire and to a lesser extent, temperature (PC2, which is forest and savanna can both occur in areas with similar climates, earlier identified via remote sensing products at the global (3, 5) and landscape (7) related to altitude) were also important drivers. Herbivory (in- scales]. In addition, for a selected area around the Gulf of Guinea with good cluding livestock and wild herbivore biomass) and sandy soils paleoecological data, we tested for historical biome transitions to confirm were found to have an equivalent and moderate impact, at this forest–savanna bistability. Second, we used tree species lists for the same continental scale. The environmental model notably predicts a forest and savanna sites to test the biome specialization of the tree species smaller forest area than the interpolated species distribution and to evaluate the overall distinctiveness of forest and savanna tree floras

Aleman et al. PNAS Latest Articles | 5of8 Downloaded by guest on September 27, 2021 A B

Ordinary Kriging Random Forest

PC1 (prec) 0.15 fire

herbivory 0.10 PC2 (temp)

Semivariance 0.05 sand CEC

500 1000 1500 2000 40 60 80 100 120 Distance (km) %IncMSE

Fig. 4. Spatial and environmental predictions of the biome index. We predicted the distribution of the biome index across the climatic space covered by the floristic surveys using (A) spatial information only vs. (B) a random forest approach based on environmental determinants. For the spatial predictions, we interpolated the distribution of the percentage of forest specialists, generalists, and savanna specialists using ordinary kriging and recomputed the biome index. The empirical semivariogram (points) and the spherical semivariogram model (lines) used for kriging and shown in A, Inset indicate that the spatial structure of generalist species is weaker and more homogeneous than that of forest and savanna specialists. For the environmental predictions of the biome index, we used climate (described by PC1 and PC2) (Fig. 1A), fire, herbivory (total biomass of both livestock and wild herbivores), and soil factors (described by the percentage of sand and cation exchange capacity, CEC, in the first 0 to 5 cm). The relative importance of each environmental determinant is shown in B, Inset. The importance (percentage increase in mean squared error, MSE) tests how the accuracy of the results is affected if the input variable is randomly permuted. White pixels in A correspond to areas outside the geographical extent and the environmental range covered by the floristic surveys. White pixels in B additionally contain pixels for which herbivory data were not available (coastal edge and edge of inland water bodies). Major water bodies are shown in blue.

(22). Third, we gridded the georeferenced herbarium records available for rainbio/index.html). These represent the largest-ever collation of georefer- tropical Africa (24, 25) and tested the degree of biome specialization char- enced herbarium records for tropical Africa (24, 25). acterizing each pixel, thereby testing the forest–savanna bistability from a We additionally gathered information on past vegetation for 42 paleo- floristic perspective. Finally, we used spatial and environmental predictions ecological sites located around the Gulf of Guinea (Dataset S2) from pub- to delineate the distribution of the forest and savanna biomes at the con- lished records of lacustrine fossil pollen (n = 24), lacustrine phytoliths (n = 1), tinental scale and to test the relative influence of climate, fire, herbivory, and phytoliths (n = 4) and δ13C of soil organic matter (n = 14) from soil and soils on the forest and savanna distributions (2, 30). profiles. We selected records with at least 2,000 y of vegetation history. From the original studies, we identified 20 paleosites with stable vegetation Datasets. during the Holocene, either forest or savanna, and 22 that experienced a Floristic data. We combined native tree species lists for 455 forest (26) and 298 biome transition, always from forest to savanna. Two sites, Lakes Bosumtwi ∼ savanna sites (27) across sub-Saharan Africa (Fig. 1D and Dataset S1). Each and Barombi Mbo, date back to the Last Glacial Maximum ( 21,000 y B.P.), site consists of a species list assigned to either forest (excluding “montane during which the vegetation was probably a mosaic of savanna (identified forest” and “thicket”) or savanna (including “woodland”) by the original by an increase in Poaceae) and forest elements. Environmental drivers. authors or by the data contributors who conducted the floristic surveys. Our study area corresponded to the geographical extent covered by the floristic surveys but excluding two extreme sites in the very Forest sites corresponded to moist broadleaved forest, although restricted to south (Fort Hare, South Africa) and east (Ogo, Somalia), ending up with lowland and terra firme, while savanna sites included both wooded grass- latitudes ranging from 28°S to 17.5°N and longitudes from 17°W to 42.5°E. lands and woodlands. Indeed, several southern African woodland types, To characterize the climatic conditions across the study area, we used the such as Miombo, mopane, and Baikiaea woodlands (28), are considered as bioclimatic variables from Worldclim version 2 at 0.5° resolution (46), which savannas since they burn regularly and have a relatively continuous grass consist of 19 variables describing precipitation and temperature, including layer (1, 20). Inclusion of a species list corresponding to either forest or sa- mean annual temperature, mean diurnal range, isothermality, temperature vanna was first based on the definition given in the source according to seasonality, maximum temperature of the warmest month, minimum tem- vegetation structure (savannas do not show closed canopy and have a grass perature of the coldest month, temperature annual range, mean tempera- layer) and/or on clear indications of species composition and ecosystem ture of the wettest quarter, mean temperature of the driest quarter, mean functioning. The combined floristic dataset analyzed consisted of a total of temperature of the warmest quarter, mean temperature of the coldest 1,707 species occurring in more than five of our combined forest and sa- quarter, MAP, precipitation of the wettest month, precipitation of the driest vanna sites, belonging to 590 genera and 110 families (Dataset S1). The month, precipitation seasonality, precipitation of the wettest quarter, pre- majority of the species were trees, but some shrubs were included in the cipitation of the driest quarter, precipitation of the warmest quarter, and savanna sites. The taxonomy was standardized according to the African Plant precipitation of the coldest quarter. = Database (www.ville-ge.ch/musinfo/bd/cjb/africa/recherche.php?langue an)in Beyond climate, we also considered disturbance, specifically fire and January 2018. chronic herbivory (herbivore biomass), and soils (sand percentage and cation To spatially extrapolate the floristic information from our forest and sa- exchange capacity) because they have been identified as determinants of vanna sites, we used independent data of plant species occurrence available savanna distribution (2, 30). Estimates of fire frequency were derived from online and assembled in the RAINBIO project (https://gdauby.github.io/ the burned area product from MODIS data at 1-km resolution (47) over the

6of8 | www.pnas.org/cgi/doi/10.1073/pnas.2011515117 Aleman et al. Downloaded by guest on September 27, 2021 2003 to 2012 period. Current day estimates of herbivore biomass at 0.5° known to be richness independent. We then performed the IndVal proce- resolution across the African continent included livestock and also historical dure (50) on the presence matrix of the 1,707 native tree species encoun- wild herbivore biomass filtered by landscape change indices (48). In this map, tered in the 455 forest sites and the 298 savanna sites. This allowed us to originally developed at 1° resolution excluding grid cells with >50% in in- identify significant indicators for categorizing species as forest or savanna land water bodies and then later refined at 0.5° resolution, there is no in- specialists or generalists (= no significant indicators) (Dataset S3). We further formation for a large edge along the coast and around inland water bodies. applied this classification of our 1,707 tree species on an independent Among available soil variables at 250-m resolution for sub-Saharan Africa dataset of species occurrence derived from georeferenced herbarium re- (https://soilgrids.org/), we retained the percentage of sand and the cation cords (24, 25). We computed the number (and percentage) of forest spe- exchange capacity in the top layer (0 to 5 cm) of soil (49). Soil variables were cialists, generalists, and savanna specialists per pixel of the 0.5°-resolution first projected and resampled to fit the 0.5° resolution. grid representing our study area (Fig. 2). The Global Lakes and Wetlands Database (GLWD; https://www.worldwildlife. We verified the concordance between the potential biomes derived from org/pages/global-lakes-and-wetlands-database) was used to map major water climate and the species composition in terms of percentage of forest spe- bodies, including lakes (GLWD-1) and rivers (GLWD-2). cialists, generalists, and savanna specialists (SI Appendix, Fig. S5). We used Kruskal–Wallis tests to compare each pair of potential biomes (forest and Data Analyses. bistable forest, bistable savanna and savanna) and specifically test whether Climatic envelope of forest and savanna. To detect the major climatic gradients the bistable region is more composed of generalist species or is a mix of over the entire study area, we performed a PCA on the matrix of climate forest and savanna specialist species. variables at 0.5° resolution (Fig. 1 and SI Appendix, Fig. S1). On the factorial Spatial distribution of the biome index. The spatial distributions of our forest plane defined by the first two principal components (PC1 and PC2), the specialists, generalists, and savanna specialists were then integrated into a pixels containing floristic surveys were colored, and the climatic space they biome index at the scale of 0.5° pixel (Fig. 3), with the following formula: cover was defined by a convex hull (Fig. 1A). The first climatic gradient (PC1, describing 42.3% of the total variance) is characterized primarily by varia- #sp for − #sp sav tion in precipitation and by seasonality of precipitation and of temperature biome index = , #sp for + #sp sav + #sp gen (Fig. 1A and SI Appendix, Fig. S1), and it differentiates forest and savanna sites. The second climatic gradient (PC2, 31.1% of variance) is driven by where #sp for is the number of forest specialists, #sp sav is the number of temperature (Fig. 1A) and differentiates northern and western savannas savanna specialists, and #sp gen is the number of generalist species in each from southern and eastern savannas and woodlands (SI Appendix, Fig. S1C), pixel. The biome index tracks the specialization of each pixel, with values which are generally found at higher altitudes (28) and thereby, exposed to toward −1 representing the dominance of savanna specialists and values colder climates (27), except for coastal plains (e.g., in Mozambique). toward +1 representing the dominance of forest specialists. Tropical forest and savanna have been shown to co-occur within the same To further examine the underlying composition of the biome index and to MAP range (3, 5, 6), which was also apparent along PC1, the more complex specifically test whether the intermediate values of the biome index rather climatic gradient integrating precipitation and seasonality (Fig. 1 A and C). ENVIRONMENTAL SCIENCES correspond to more generalist species or to a mix of forest and savanna We defined this climatic space as the region of potential bistability, or specialists, we used spline regressions between the percentage of each of the bistable region, because forest and savanna are both widespread and be- three species groups and the biome index (SI Appendix, Fig. S6). Since the cause the stability over time has been previously demonstrated through frequency distribution of the biome index was found to be slightly trimodal mechanistic modeling (6). Because of the nonuniform distribution of sites (Fig. 3, Inset), we mapped the three states (the savanna state, the interme- along PC1, we subsampled the sites with a stratified random sampling by taking, with replacement, 20 sites for each bin of 0.1 units PC1 (i.e., 20 diate state, and the forest states) using different thresholds on the biome samples with PC1 between −4 and −3.9 and so on). Numerically, the bistable index (SI Appendix, Fig. S7). Finally, to examine the effect of herbarium region was defined as the region over which the slope in the frequency of record availability, we constructed the map of the biome index for pixels the two biomes at each point over PC1 exceeded the mean of the slope containing at least 5 or 10 of our species and showed the associated fre- averaged over the entire range (SI Appendix, Fig. S2A). In practice, we quency distribution (SI Appendix, Fig. S8). plotted the frequency of forest and savanna sites along PC1 and computed Spatial and environmental predictions of the biome index. Because species oc- the slope of these curves for each PC1 bin. The slopes were averaged for the currences were missing in remote and/or undersampled areas (Fig. 3), we entire PC1 range. The bistable region was then defined as the area of the recomputed our biome index (Fig. 4A) from spatially interpolated values of plot for which individual slopes are above the average. We found that forest the percentage of forest specialists, generalists, and savanna specialists using and savanna sites coexist between PC1 values of 0.141 and 4.235 (Fig. 1C) ordinary kriging (SI Appendix, Fig. S9 A–C). The spherical autocorrelation and that their frequency distribution is equal to 0.5 for a PC1 value of 1.943, function provided a good fit to the experimental variograms (Fig. 4 A, Inset). such that below this value, the probability of savanna occurring is higher We evaluated the spatial accuracy of the kriged map outside the pixels than the probability of forest and vice versa. containing herbarium records of our study species by computing CIs from After defining these thresholds, we mapped the potential biomes based the variance of the predictions (SI Appendix, Fig. S9 D–F). on gridded climatic variables, but we restricted our spatial predictions to the We also used a random forest approach to identify the environmental geographic extent and to the climate space covered by the floristic surveys. determinants of the biome index and to map the biome index according to The latter was defined with a convex hull on the site scores along PC1 and PC2 climate (described by PC1 and PC2) (Fig. 1A and SI Appendix, Fig. S1), fire (SI (Fig. 1A). Predictions were thus not possible for some mountain areas in Appendix, Fig. S3), herbivory, and soil factors (described by the percentage eastern Africa and for a large area in southern Africa corresponding to of sand and cation exchange capacity, CEC, in the top 0 to 5 cm). For the (semidesert) grasslands (28). First, savanna is defined as having values of random forest, our dataset corresponded to the pixels containing at least 5 < ≥ PC1 0.141, and forest is defined as having PC1 4.235 (Fig. 1C and SI of our 1,707 species and with available data for all environmental deter- Appendix, Fig. S2A). Then, within the bistable region, bistable savanna, minants. We kept 10% of the dataset for validation (n = 169 pixels) and fit which corresponds to a greater probability of savanna in the climatic space the random forest model on a calibration dataset (n = 1,523 pixels). We also of co-occurrence (Fig. 1C), is defined by 0.141 ≤ PC1 < 1.943, and bistable examined the importance of each environmental determinant, which mea- forest, which corresponds to a greater probability of forest in the climatic sures how the accuracy of the results is affected if the input variable is space of co-occurrence, is defined by 1.943 ≤ PC1 < 4.235. Finally, for clarity randomly permuted. and comparison with earlier results, we conducted the same analysis with MAP, sampling 20 sites for each MAP bin of 50 mm (SI Appendix, Fig. S2B), All of the analyses were performed within the open source R environment and found a range of 700 to 1,900 mm for the bistable region (SI Appendix, (51) using the raster package (52) for raster constructions and most spatial Fig. S2C). analyses. The ade4 package (53) was used for the PCA, and the dismo package To confirm the forest–savanna bistability, we mapped the paleoecological (54) was used for the convex hull. The vegan package (55) was used for the evidence of biome transitions (Dataset S2). To specifically test whether the dissimilarity analysis, and the labdsv package (56) was used for computing bistable region is more likely to burn than the other potential biomes, we species indicator values (and significance) for each biome. The gstat package computed mean fire frequency for each potential biome (SI Appendix, Fig. (57) was used for the spatial interpolation by ordinary kriging, while the Ran- S3). The latter analysis was performed both at the pixel and at the site scale. domForest package (58) for the Random Forest regression was used to identify Floristics of forest and savanna. We first computed the overall dissimilarity in the environmental determinants and to provide environmental predictions. tree species composition between all pairs of forest and savanna sites (SI

Appendix, Fig. S4) using the Simpson index of β diversity (βsim), which is Data Availability. All study data are included in the article and SI Appendix.

Aleman et al. PNAS Latest Articles | 7of8 Downloaded by guest on September 27, 2021 ACKNOWLEDGMENTS. We thank all the members of the Tropical Forestry 143/A3/HERBAXYLAREDD (to A.F.) and Natural Environment Research Coun- group in Gembloux Agro-Bio Tech, University of Liège, not listed as coau- cil Grant NE/P008755/1 (to K.G.D. and C.M.R.). D.B. received support from thors of the study. This work was funded by Belgian Science Policy Grant BR/ Fondation Wiener-Anspach.

1. J. Ratnam et al., When is a ‘forest’ a savanna, and why does it matter? Glob. Ecol. 30. M. Sankaran et al., Determinants of woody cover in African savannas. Nature 438, Biogeogr. 20, 653–660 (2011). 846–849 (2005). 2. C. E. Lehmann, S. A. Archibald, W. A. Hoffmann, W. J. Bond, Deciphering the distri- 31. U. Salzmann, P. Hoelzmann, The Dahomey gap: An abrupt climatically induced rain bution of the savanna biome. New Phytol. 191, 197–209 (2011). forest fragmentation in West Africa during the late Holocene. Holocene 15, 190–199 3. A. C. Staver, S. Archibald, S. A. Levin, The global extent and determinants of savanna (2005). and forest as alternative biome states. Science 334, 230–232 (2011). 32. J. C. Aleman et al., Palaeo-trajectories of forest savannization in the southern Congo. 4. P. L. Silva de Miranda et al., Using tree species inventories to map biomes and assess Biol. Lett. 15, 20190284 (2019). their climatic overlaps in lowland tropical South America. Glob. Ecol. Biogeogr. 27, 33. U. Salzmann, Are modern savannas degraded forests?-A Holocene pollen record from 899–912 (2018). the Sudanian vagetation zone of NE Nigeria. Veg. Hist. Archaeobot. 9,1–15 (2000). 5. M. Hirota, M. Holmgren, E. H. Van Nes, M. Scheffer, Global resilience of tropical forest 34. R. R. Klopper, L. Gautier, C. Chatelain, G. F. Smith, R. Spichiger, Floristics of the an- and savanna to critical transitions. Science 334, 232–235 (2011). giosperm flora of sub-Saharan Africa: An analysis of the African plant checklist and 6. A. C. Staver, S. Archibald, S. Levin, Tree cover in sub-Saharan Africa: Rainfall and fire Database. Taxon 56, 201–208 (2007). constrain forest and savanna as alternative stable states. Ecology 92, 1063–1072 35. J. C. Aleman, M. A. Jarzyna, A. C. Staver, Forest extent and deforestation in tropical (2011). Africa since 1900. Nat. Ecol. Evol. 2,26–33 (2018). 7. C. Favier et al., Abrupt shifts in African savanna tree cover along a climatic gradient. 36. A. Vincens, Y. Garcin, G. Buchet, Influence of rainfall seasonality on African lowland Glob. Ecol. Biogeogr. 21, 787–797 (2012). vegetation during the late quaternary: Pollen evidence from Lake Masoko, Tanzania. 8. G. P. Hempson, S. Archibald, J. E. Donaldson, C. E. R. Lehmann, Alternate grassy J. Biogeogr. 34, 1274–1288 (2007). ecosystem states are determined by palatability-flammability trade-offs. Trends Ecol. 37. L. Laestadius et al., Opportunities for forest landscape restoration. Unasylva 62, 238 Evol. 34, 286–290 (2019). (2011). 9. W. A. Hoffmann et al., Ecological thresholds at the savanna-forest boundary: How 38. M. Greve, B. Reyers, A. Mette Lykke, J.-C. Svenning, Spatial optimization of carbon- plant traits, resources and fire govern the distribution of tropical biomes. Ecol. Lett. stocking projects across Africa integrating stocking potential with co-benefits and 15, 759–768 (2012). feasibility. Nat. Commun. 4, 2975 (2013). 10. T. Charles-Dominique, A. C. Staver, G. F. Midgley, W. J. Bond, Functional differenti- 39. R. C. R. Abreu et al., The biodiversity cost of carbon sequestration in tropical savanna. ation of biomes in an African savanna/forest mosaic. S. Afr. J. Bot. 101,82–90 (2015). Sci. Adv. 3, e1701284 (2017). 11. M. D. Swaine, J. B. Hall, J. M. Lock, The forest-savanna boundary in west-central 40. B. P. Murphy, A. N. Andersen, C. L. Parr, The underestimated biodiversity of tropical Ghana. Ghana J. Sci. 16,35–52 (1976). grassy biomes. Philos. Trans. R Soc. Lond. B Biol. Sci. 371, 20150319 (2016). 12. J. C. Aleman, A. C. Staver, Spatial patterns in the global distributions of savanna and 41. C. M. Ryan et al., Ecosystem services from southern African woodlands and their fu- forest. Glob. Ecol. Biogeogr. 27, 792–803 (2018). ture under global change. Philos. Trans. R Soc. Lond. B Biol. Sci. 371, 20150312 (2016). 13. C. Xu et al., Remotely sensed canopy height reveals three pantropical ecosystem 42. N. Stevens, C. E. Lehmann, B. P. Murphy, G. Durigan, Savanna woody encroachment is states. Ecology 97, 2518–2521 (2016). widespread across three continents. Glob. Change Biol. 23, 235–244 (2017). 14. V. de L. Dantas, M. Hirota, R. S. Oliveira, J. G. Pausas, Disturbance maintains alter- 43. E. Buisson et al., Resilience and restoration of tropical and subtropical grasslands, native biome states. Ecol. Lett. 19,12–19 (2016). savannas, and grassy woodlands. Biol. Rev. Camb. Philos. Soc. 94, 590–609 (2019). 15. M. C. Hansen et al., Towards an operational MODIS continuous field of percent tree 44. A. T. Kaptué, L. Prihodko, N. P. Hanan, On regreening and degradation in Sahelian cover algorithm: Examples using AVHRR and MODIS data. Remote Sens. Environ. 83, watersheds. Proc. Natl. Acad. Sci. U.S.A. 112, 12133–12138 (2015). 303–319 (2002). 45. L. Gillson, R. Marchant, From myopia to clarity: Sharpening the focus of ecosystem 16. M. C. Hansen et al., High-resolution global maps of 21st-century forest cover change. management through the lens of palaeoecology. Trends Ecol. Evol. 29, 317–325 Science 342, 850–853 (2013). (2014). 17. J. C. Aleman, O. Blarquez, C. A. Staver, Land-use change outweighs projected effects 46. S. E. Fick, R. J. Hijmans, WorldClim 2: New 1-km spatial resolution climate surfaces for of changing rainfall on tree cover in sub-Saharan Africa. Glob. Change Biol. 22, global land areas. Int. J. Climatol. 37, 4302–4315 (2017). 3013–3025 (2016). 47. L. Giglio, J. Descloitres, C. O. Justice, Y. J. Kaufman, An enhanced contextual fire 18. M. Torello-Raventos et al., On the delineation of tropical vegetation types with an detection algorithm for MODIS. Remote Sens. Environ. 87, 273–282 (2003). emphasis on forest/savanna transitions. Plant Ecol. Divers. 6, 101–137 (2013). 48. G. P. Hempson, S. Archibald, W. J. Bond, The consequences of replacing wildlife with 19. J. W. Veldman, Clarifying the confusion: Old-growth savannahs and tropical ecosys- livestock in Africa. Sci. Rep. 7, 17196 (2017). tem degradation. Philos. Trans. R Soc. Lond. B Biol. Sci. 371, 20150306 (2016). 49. T. Hengl et al., Mapping soil properties of Africa at 250 m resolution: Random forests 20. C. P. Osborne et al., Human impacts in African savannas are mediated by plant significantly improve current predictions. PLoS One 10, e0125814 (2015). functional traits. New Phytol. 220,10–24 (2018). 50. M. Dufrêne, P. Legendre, Species assemblages and indicator species: The need for a 21. H. P. Linder et al., The partitioning of Africa: Statistically defined biogeographical flexible asymmetrical approach. Ecol. Monogr. 67, 345–366 (1997). regions in sub‐Saharan Africa. J. Biogeogr. 39, 1189–1205 (2012). 51. R Development Core Team, R: A Language and Environment for Statistical Comput- 22. H. P. Linder, The evolution of African plant diversity. Front. Ecol. Evol. 2, 38 (2014). ing, Version 3.5.0. https://www.R-project.org/. Accessed 23 April 2018. 23. G. P. Hempson, S. Archibald, W. J. Bond, A continent-wide assessment of the form and 52. R. J. Hijmans, raster: Geographic Data Analysis and Modeling, R package Version 2.6- intensity of large mammal herbivory in Africa. Science 350, 1056–1061 (2015). 7. https://CRAN.R-project.org/package=raster. Accessed 14 June 2018. 24. V. Droissart et al., Beyond trees: Biogeographical regionalization of tropical Africa. 53. S. Dray, A. B. Dufour, The ade4 package: Implementing the duality diagram for J. Biogeogr. 45, 1153–1167 (2018). ecologists. J. Stat. Softw. 22,1–20 (2007). 25. M. S. Sosef et al., Exploring the floristic diversity of tropical Africa. BMC Biol. 15,15 54. R. J. Hijmans, S. Phillips, J. Leathwick, J. Elith, dismo: Species Distribution Modeling, R (2017). package Version 1.1-4. https://CRAN.R-project.org/package=dismo. Accessed 14 June 26. A. Fayolle et al., Patterns of tree species composition across tropical African forests. 2018. J. Biogeogr. 41, 2320–2331 (2014). 55. J. Oksanen et al, vegan: Community Ecology Package, R package Version 2.5-2. https:// 27. A. Fayolle et al., A sharp floristic discontinuity revealed by the biogeographic re- CRAN.R-project.org/package=vegan. Accessed 14 June 2018. gionalization of African savannas. J. Biogeogr. 46, 454–465 (2019). 56. D. W. Roberts, labdsv: Ordination and Multivariate Analysis for Ecology, R package 28. F. White, The Vegetation of Africa: A Descriptive Memoir to Accompany the UNESCO/ Version 1.8-0. https://CRAN.R-project.org/package=labdsv. Accessed 14 June 2018. AETFAT/UNSO Vegetation Map of Africa (ORSTOM–UNESCO, 1983). 57. E. J. Pebesma, Multivariable geostatistics in {S}: The gstat package. Comput. Geosci. 29. A. Vincens et al., Forest response to climate changes in Atlantic Equatorial Africa 30, 683–691 (2004). during the last 4000 years BP and inheritance on the modern landscapes. J. Biogeogr. 58. A. Liaw, M. Wiener, Classification and regression by randomForest. R News 2, 26, 879–885 (1999). 18–22 (2002).

8of8 | www.pnas.org/cgi/doi/10.1073/pnas.2011515117 Aleman et al. Downloaded by guest on September 27, 2021